one thread / before → after
The same voice, on either side of what I said about my state or care. Each connection keeps its own date.
Each thread is one retained reference group with a single alignment date and at least five eligible days on each side. The left semicircle shows the before median; the right shows the after median. On both sides, values increase from bottom to top. Each thread connects its two measured values; its bend is a visual construction. The scale expands the observed minimum-to-maximum range across all eligible groups and both full and held-out measures. Counts and words use log(1 + value) within that range; shares use a linear observed range. Zero may lie outside the displayed range: the lower and upper bounds are labeled, and exact values are unchanged. Category filters and holding out health posts keep the same scale. Red triangles mark high/mixed references, blue circles mark low references, and yellow diamonds mark care references; these categories describe the references, not the surrounding days. The 7 / 14 / 21-day control changes the comparison window.
Exact dates align to the stated date; current references align to the report date. Groups with uncertain recent/month/year intervals are omitted from this point-aligned artwork rather than given a precise event day. Report dates and group IDs remain in the reader. Repeated references with the same category and date bounds were consolidated. Holding out health-coded posts tests sensitivity to reusing label-defining records; it does not create independent clinical labels. Messages sent and received are a separate read-only snapshot, distinct from the older exported headers, and the hold-out control does not apply to them. Windows overlap and reference selection is not independent. Group counts are not episode counts.
Full collection method · Nine forms